Fix per file ruff ignores related to simplification (#26259)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
@@ -439,10 +439,7 @@ def mean_dim(
|
||||
output = torch.empty(output_shape, dtype=dtype, device=input.device)
|
||||
|
||||
# Reshape output for kernel
|
||||
if keepdim:
|
||||
output_2d = output.reshape(M, 1, K).squeeze(1)
|
||||
else:
|
||||
output_2d = output.reshape(M, K)
|
||||
output_2d = output.reshape(M, 1, K).squeeze(1) if keepdim else output.reshape(M, K)
|
||||
|
||||
# Launch kernel
|
||||
grid = (M * K,)
|
||||
|
||||
@@ -151,10 +151,7 @@ def chunk_fwd_o(
|
||||
) -> torch.Tensor:
|
||||
B, T, Hg, K, V = *q.shape, v.shape[-1]
|
||||
H = v.shape[-2]
|
||||
if FLA_GDN_FIX_BT:
|
||||
BT = 64
|
||||
else:
|
||||
BT = min(chunk_size, max(16, triton.next_power_of_2(T)))
|
||||
BT = 64 if FLA_GDN_FIX_BT else min(chunk_size, max(16, triton.next_power_of_2(T)))
|
||||
chunk_indices = (
|
||||
prepare_chunk_indices(cu_seqlens, BT) if cu_seqlens is not None else None
|
||||
)
|
||||
|
||||
@@ -1746,10 +1746,7 @@ def fused_experts_impl(
|
||||
else:
|
||||
raise ValueError(f"Unsupported compute_type: {hidden_states.dtype}")
|
||||
|
||||
if inplace:
|
||||
out_hidden_states = hidden_states
|
||||
else:
|
||||
out_hidden_states = torch.empty_like(hidden_states)
|
||||
out_hidden_states = hidden_states if inplace else torch.empty_like(hidden_states)
|
||||
|
||||
if use_mxfp4_w4a4:
|
||||
# Weight has to be dequantized for mxfp4 emulation.
|
||||
|
||||
@@ -886,10 +886,7 @@ def determine_expert_map(
|
||||
# Distribute experts as evenly as possible to each rank.
|
||||
base_experts = global_num_experts // ep_size
|
||||
remainder = global_num_experts % ep_size
|
||||
if ep_rank < remainder:
|
||||
local_num_experts = base_experts + 1
|
||||
else:
|
||||
local_num_experts = base_experts
|
||||
local_num_experts = base_experts + 1 if ep_rank < remainder else base_experts
|
||||
|
||||
# Create a tensor of size num_experts filled with -1
|
||||
expert_map = torch.full((global_num_experts,), -1, dtype=torch.int32)
|
||||
|
||||
@@ -948,10 +948,7 @@ class FusedMoEModularKernel(torch.nn.Module):
|
||||
"""
|
||||
|
||||
a1 = hidden_states
|
||||
if inplace and self.shared_experts is None:
|
||||
output = a1
|
||||
else:
|
||||
output = torch.zeros_like(a1)
|
||||
output = a1 if inplace and self.shared_experts is None else torch.zeros_like(a1)
|
||||
|
||||
local_num_experts = w1.size(0)
|
||||
if global_num_experts == -1:
|
||||
|
||||
@@ -355,10 +355,7 @@ def rocm_aiter_fused_experts(
|
||||
topk_weights = topk_weights.to(torch.float32)
|
||||
topk_ids = topk_ids.to(torch.int32)
|
||||
|
||||
if expert_map is not None:
|
||||
expert_mask = (expert_map > -1).to(torch.int32)
|
||||
else:
|
||||
expert_mask = None
|
||||
expert_mask = (expert_map > -1).to(torch.int32) if expert_map is not None else None
|
||||
|
||||
# w8a8 per-channel quantization
|
||||
if (
|
||||
|
||||
@@ -318,10 +318,7 @@ class GemmaRMSNorm(CustomOp):
|
||||
"""PyTorch-native implementation equivalent to forward()."""
|
||||
orig_dtype = x.dtype
|
||||
if residual is not None:
|
||||
if orig_dtype == torch.float16:
|
||||
x = x + residual.float()
|
||||
else:
|
||||
x = x + residual
|
||||
x = x + residual.float() if orig_dtype == torch.float16 else x + residual
|
||||
residual = x
|
||||
|
||||
x = x.float()
|
||||
|
||||
@@ -207,10 +207,7 @@ def _fwd_kv_parallel(
|
||||
kv = tl.zeros([D_FBLOCK, E_FBLOCK], dtype=tl.float32)
|
||||
|
||||
# Handle the last block which might be smaller than BLOCK
|
||||
if off_block == NUM_BLOCK - 1:
|
||||
split_n = n - (NUM_BLOCK - 1) * BLOCK
|
||||
else:
|
||||
split_n = BLOCK
|
||||
split_n = n - (NUM_BLOCK - 1) * BLOCK if off_block == NUM_BLOCK - 1 else BLOCK
|
||||
left_shift = tl.cdiv(split_n, CBLOCK) * CBLOCK - split_n
|
||||
num_blocks = min(tl.cdiv(split_n, CBLOCK), NUM_CBLOCK)
|
||||
k_decay_ptr += (NUM_CBLOCK - num_blocks) * CBLOCK
|
||||
|
||||
@@ -502,15 +502,11 @@ class CompressedTensorsConfig(QuantizationConfig):
|
||||
QuantizationStrategy.CHANNEL,
|
||||
QuantizationStrategy.BLOCK,
|
||||
]
|
||||
if not (
|
||||
return (
|
||||
is_symmetric_weight
|
||||
and is_static_weight # noqa: SIM103
|
||||
and is_static_weight
|
||||
and is_tensor_or_channel_or_block_weight
|
||||
):
|
||||
return False
|
||||
|
||||
# All conditions satisfied.
|
||||
return True
|
||||
)
|
||||
|
||||
def _is_wNa16_group_channel(
|
||||
self, weight_quant: QuantizationArgs, input_quant: QuantizationArgs
|
||||
|
||||
@@ -80,10 +80,7 @@ def check_equal_or_regex_match(layer_name: str, targets: Iterable[str]) -> bool:
|
||||
Checks whether a layer_name is exactly equal or a regex match for
|
||||
if target starts with 're:' to any target in list.
|
||||
"""
|
||||
for target in targets:
|
||||
if _is_equal_or_regex_match(layer_name, target):
|
||||
return True
|
||||
return False
|
||||
return any(_is_equal_or_regex_match(layer_name, target) for target in targets)
|
||||
|
||||
|
||||
def find_matched_target(
|
||||
|
||||
@@ -81,10 +81,7 @@ def check_equal_or_regex_match(layer_name: str, targets: Iterable[str]) -> bool:
|
||||
Checks whether a layer_name is exactly equal or a regex match for
|
||||
if target starts with 're:' to any target in list.
|
||||
"""
|
||||
for target in targets:
|
||||
if _is_equal_or_regex_match(layer_name, target):
|
||||
return True
|
||||
return False
|
||||
return any(_is_equal_or_regex_match(layer_name, target) for target in targets)
|
||||
|
||||
|
||||
def _is_equal_or_regex_match(
|
||||
|
||||
Reference in New Issue
Block a user